Attention hierarchical network for super-resolution

Deep neural networks with attention mechanism for super-resolution (SR) have achieved good SR performance by focusing on the high-frequency components of images. However, during the SR process, it is difficult for these networks to obtain multi-level high-frequency features with different extraction difficulties from low-resolution images, resulting in the lack of textures and details in the reconstructed SR images. To solve this problem, we propose an attention hierarchical network (AHN) for SR. The proposed AHN separates and extracts high-frequency features with different extraction difficulties in a hierarchical way to obtain multi-level high-frequency features. In the process of separation and extraction, we separate high-frequency features into easy-to-extract features and difficult-to-extract features by attention block and extract the separated features by dense-residual module. Extensive experiments demonstrate that the proposed AHN is superior to the state-of-the-art SR methods and reconstructs better SR images that contain more textures and details.